Fast and accurate detection of lactating sow nursing behavior with CNN-based optical flow and features

Abstract Automated detection of sow nursing behavior is beneficial to the health, welfare, and productivity of sows and piglets in the commercial swine industry. We proposed a fast and accurate detection method for sow nursing behavior using CNN-based optical flow and features. The detection method has two major steps. In the first step, a spatial localization network consisting of a sow detector and a keypoint detector was developed in order to localize the nursing zone. In the second step, a nursing behavior recognition network was developed to detect sow nursing behaviors from untrimmed videos. To achieve this, a CNN-based optical flow estimator was used to produce an optical flow within the nursing zone. RGB images of the detected nursing zone and the corresponding optical flow images were then fed into the recognition network in sequence in order to capture spatiotemporal information. In our study, four separate convolutional neural networks were trained. 11.7 h of videos collected from 19 pens in sow farms were used to train the model, and 24.7 h of other videos from 8 pens were tested for validation. The results revealed that the proposed method achieved an accuracy of 97.63%, a sensitivity of 98.55%, a specificity of 97.36% and a detection speed of 6.36 fps from untrimmed videos. Compared with our previous method, this method has a superior performance in terms of accuracy and sensitivity. In addition, this method has a 7.45% smaller model size, yet with a 176.52% faster detection speed, which demonstrates the feasibility of automated monitoring of nursing behavior.